From the source
Hugging Face published an explainer on decision models, which take a state and a set of predefined answers and return a structured decision with probabilities instead of generating text.
- The post contrasts them with prompting an LLM for JSON, which can produce invalid output, unexpected keys or answers outside the allowed choices, and with constrained decoding libraries such as Outlines and XGrammar that enforce schemas but add computational work.
- It traces the term to
TypeSafe AI's Jev, described as a System One model, and says the community settled on "decision model" because it names the model's computational role rather than a theory of cognition.
- The interface is given as state plus questions plus allowed answers producing decisions and probabilities, illustrated with a customer-support ticket routed to billing, technical or sales.
- The post also compares decision models with traditional classification, whose output classes are baked into the classification head, and with zero-shot classification, which takes candidate labels at inference time.





